gradient share problem

[vergilus]

in pytorch 1.0 the following line:

https://github.com/ikostrikov/pytorch-a3c/blob/master/train.py#L14

returns the gradient assignment error:
RuntimeError: assigned grad has data of a different type.

any fixes or suggestions?

[vergilus]

I think I figure it out.
there is no grad in global params (None type), the way to solve this problem is to move the local parameters to CPU, then you can assign local parameter. to make life easier, you can wrap this function in the model.
ensure_shared_grads(model, shared_model)
becomes
model.to("cpu").ensure_shared_grads(shared_model)
but do mind that when you run your model on GPU, you have to switch it back to the device which it was running for next backward() (in case sometimes model can have "expected GPU but got CPU backend" error)